mllm_encoder.py

# coding=utf-8
# Copyright 2025 The Qwen Team and The HuggingFace Inc. team. All rights reserved.
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import base64
import copy
import math
import os
import sys
import time
import warnings
from concurrent.futures import ThreadPoolExecutor
from dataclasses import dataclass
from functools import lru_cache
from io import BytesIO
from typing import Any, Callable, Dict, List, Optional, Tuple, Union

import numpy as np
import requests
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision
from packaging import version
from PIL import Image
from torchvision import io, transforms
from torchvision.transforms import InterpolationMode

from diffusers import ModelMixin, ConfigMixin
from diffusers.configuration_utils import register_to_config
from transformers.activations import ACT2FN
from transformers.cache_utils import Cache, DynamicCache
from transformers.generation import GenerationMixin
from transformers.masking_utils import create_causal_mask, create_sliding_window_causal_mask
from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
from transformers.modeling_layers import GradientCheckpointingLayer
from transformers.modeling_outputs import BaseModelOutputWithPast, ModelOutput
from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from transformers.processing_utils import Unpack
from transformers.utils import TransformersKwargs, auto_docstring, can_return_tuple, is_torchdynamo_compiling, logging
from transformers.utils.deprecation import deprecate_kwarg
from transformers.models.qwen2.modeling_qwen2 import Qwen2RMSNorm
from transformers.models.qwen2_5_vl.configuration_qwen2_5_vl import Qwen2_5_VLConfig, Qwen2_5_VLTextConfig, Qwen2_5_VLVisionConfig

logger = logging.get_logger(__name__)

# ─────────────────────────────────────────────────────────────────────────────
# Vision processing utilities (formerly qwen_vl_utils.py)
# ─────────────────────────────────────────────────────────────────────────────

MAX_RATIO = 200
SPATIAL_MERGE_SIZE = 2
IMAGE_MIN_TOKEN_NUM = 4
IMAGE_MAX_TOKEN_NUM = 16384
VIDEO_MIN_TOKEN_NUM = 128
VIDEO_MAX_TOKEN_NUM = 768

FPS = 2.0
FRAME_FACTOR = 2
FPS_MIN_FRAMES = 4
FPS_MAX_FRAMES = 16
MAX_NUM_WORKERS_FETCH_VIDEO = 8

MODEL_SEQ_LEN = int(float(os.environ.get('MODEL_SEQ_LEN', 128000)))


# ─────────────────────────────────────────────────────────────────────────────
# Qwen2.5-VL model (formerly modeling_qwen2_5_vl.py)
# ─────────────────────────────────────────────────────────────────────────────

class Qwen2_5_VLMLP(nn.Module):
    def __init__(self, config, bias: bool = False):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size
        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=bias)
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=bias)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=bias)
        self.act_fn = ACT2FN[config.hidden_act]

    def forward(self, hidden_state):
        return self.down_proj(self.act_fn(self.gate_proj(hidden_state)) * self.up_proj(hidden_state))


class Qwen2_5_VisionPatchEmbed(nn.Module):
    def __init__(
        self,
        patch_size: int = 14,
        temporal_patch_size: int = 2,
        in_channels: int = 3,
        embed_dim: int = 1152,
    ) -> None:
        super().__init__()
        self.patch_size = patch_size
        self.temporal_patch_size = temporal_patch_size
        self.in_channels = in_channels
        self.embed_dim = embed_dim

        kernel_size = [temporal_patch_size, patch_size, patch_size]
        self.proj = nn.Conv3d(in_channels, embed_dim, kernel_size=kernel_size, stride=kernel_size, bias=False)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        target_dtype = self.proj.weight.dtype
        hidden_states = hidden_states.view(
            -1, self.in_channels, self.temporal_patch_size, self.patch_size, self.patch_size
        )
        hidden_states = self.proj(hidden_states.to(dtype=target_dtype)).view(-1, self.embed_dim)
        return hidden_states


class Qwen2_5_VisionRotaryEmbedding(nn.Module):
    inv_freq: torch.Tensor  # fix linting for `register_buffer`

    def __init__(self, dim: int, theta: float = 10000.0) -> None:
        super().__init__()
        inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float) / dim))
        self.register_buffer("inv_freq", inv_freq, persistent=False)

    def forward(self, seqlen: int) -> torch.Tensor:
        seq = torch.arange(seqlen, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
        freqs = torch.outer(seq, self.inv_freq)
        return freqs


class Qwen2_5_VLPatchMerger(nn.Module):
    def __init__(self, dim: int, context_dim: int, spatial_merge_size: int = 2) -> None:
        super().__init__()
        self.hidden_size = context_dim * (spatial_merge_size**2)
        self.ln_q = Qwen2RMSNorm(context_dim, eps=1e-6)
        self.mlp = nn.Sequential(
            nn.Linear(self.hidden_size, self.hidden_size),
            nn.GELU(),
            nn.Linear(self.hidden_size, dim),
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = self.mlp(self.ln_q(x).view(-1, self.hidden_size))
        return x


def rotate_half(x):
    """Rotates half the hidden dims of the input."""
    x1 = x[..., : x.shape[-1] // 2]
    x2 = x[..., x.shape[-1] // 2 :]
    return torch.cat((-x2, x1), dim=-1)


def apply_rotary_pos_emb_vision(
    q: torch.Tensor, k: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor
) -> tuple[torch.Tensor, torch.Tensor]:
    orig_q_dtype = q.dtype
    orig_k_dtype = k.dtype
    q, k = q.float(), k.float()
    cos, sin = cos.unsqueeze(-2).float(), sin.unsqueeze(-2).float()
    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    q_embed = q_embed.to(orig_q_dtype)
    k_embed = k_embed.to(orig_k_dtype)
    return q_embed, k_embed


def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
    """
    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
    """
    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
    if n_rep == 1:
        return hidden_states
    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)


def eager_attention_forward(
    module: nn.Module,
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    attention_mask: Optional[torch.Tensor],
    scaling: float,
    dropout: float = 0.0,
    **kwargs,
):
    key_states = repeat_kv(key, module.num_key_value_groups)
    value_states = repeat_kv(value, module.num_key_value_groups)

    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
    if attention_mask is not None:
        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
        attn_weights = attn_weights + causal_mask

    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
    attn_output = torch.matmul(attn_weights, value_states)
    attn_output = attn_output.transpose(1, 2).contiguous()

    return attn_output, attn_weights


class Qwen2_5_VLVisionAttention(nn.Module):
    def __init__(self, config: Qwen2_5_VLVisionConfig) -> None:
        super().__init__()
        self.dim = config.hidden_size
        self.num_heads = config.num_heads
        self.head_dim = self.dim // self.num_heads
        self.num_key_value_groups = 1  # needed for eager attention
        self.qkv = nn.Linear(self.dim, self.dim * 3, bias=True)
        self.proj = nn.Linear(self.dim, self.dim)
        self.scaling = self.head_dim**-0.5
        self.config = config
        self.attention_dropout = 0.0
        self.is_causal = False

    def forward(
        self,
        hidden_states: torch.Tensor,
        cu_seqlens: torch.Tensor,
        rotary_pos_emb: Optional[torch.Tensor] = None,
        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
        **kwargs,
    ) -> torch.Tensor:
        seq_length = hidden_states.shape[0]
        query_states, key_states, value_states = (
            self.qkv(hidden_states).reshape(seq_length, 3, self.num_heads, -1).permute(1, 0, 2, 3).unbind(0)
        )
        cos, sin = position_embeddings
        query_states, key_states = apply_rotary_pos_emb_vision(query_states, key_states, cos, sin)

        query_states = query_states.transpose(0, 1).unsqueeze(0)
        key_states = key_states.transpose(0, 1).unsqueeze(0)
        value_states = value_states.transpose(0, 1).unsqueeze(0)

        attention_interface: Callable = eager_attention_forward
        if self.config._attn_implementation != "eager":
            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]

        if self.config._attn_implementation == "flash_attention_2":
            # Flash Attention 2: Use cu_seqlens for variable length attention
            max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
            attn_output, _ = attention_interface(
                self,
                query_states,
                key_states,
                value_states,
                attention_mask=None,
                scaling=self.scaling,
                dropout=0.0 if not self.training else self.attention_dropout,
                cu_seq_lens_q=cu_seqlens,
                cu_seq_lens_k=cu_seqlens,
                max_length_q=max_seqlen,
                max_length_k=max_seqlen,
                is_causal=False,
                **kwargs,
            )
        else:
            # Other implementations: Process each chunk separately
            lengths = cu_seqlens[1:] - cu_seqlens[:-1]
            splits = [
                torch.split(tensor, lengths.tolist(), dim=2) for tensor in (query_states, key_states, value_states)
            ]

            attn_outputs = [
                attention_interface(
                    self,
                    q,
                    k,
                    v,
                    attention_mask=None,
                    scaling=self.scaling,
                    dropout=0.0 if not self.training else self.attention_dropout,
                    is_causal=False,
                    **kwargs,
                )[0]
                for q, k, v in zip(*splits)
            ]
            attn_output = torch.cat(attn_outputs, dim=1)

        attn_output = attn_output.reshape(seq_length, -1).contiguous()
        attn_output = self.proj(attn_output)
        return attn_output


class Qwen2_5_VLVisionBlock(GradientCheckpointingLayer):
    def __init__(self, config, attn_implementation: str = "sdpa") -> None:
        super().__init__()
        self.norm1 = Qwen2RMSNorm(config.hidden_size, eps=1e-6)
        self.norm2 = Qwen2RMSNorm(config.hidden_size, eps=1e-6)
        self.attn = Qwen2_5_VLVisionAttention(config=config)
        self.mlp = Qwen2_5_VLMLP(config, bias=True)

    def forward(
        self,
        hidden_states: torch.Tensor,
        cu_seqlens: torch.Tensor,
        rotary_pos_emb: Optional[torch.Tensor] = None,
        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
        **kwargs,
    ) -> torch.Tensor:
        hidden_states = hidden_states + self.attn(
            self.norm1(hidden_states),
            cu_seqlens=cu_seqlens,
            rotary_pos_emb=rotary_pos_emb,
            position_embeddings=position_embeddings,
            **kwargs,
        )
        hidden_states = hidden_states + self.mlp(self.norm2(hidden_states))
        return hidden_states


@auto_docstring
class Qwen2_5_VLPreTrainedModel(PreTrainedModel):
    config: Qwen2_5_VLConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["Qwen2_5_VLDecoderLayer", "Qwen2_5_VLVisionBlock"]
    _skip_keys_device_placement = "past_key_values"
    _supports_flash_attn = True
    _supports_sdpa = True

    _can_compile_fullgraph = True
    _supports_attention_backend = True


class Qwen2_5_VisionTransformerPretrainedModel(Qwen2_5_VLPreTrainedModel):
    config: Qwen2_5_VLVisionConfig
    _no_split_modules = ["Qwen2_5_VLVisionBlock"]

    def __init__(self, config, *inputs, **kwargs) -> None:
        super().__init__(config, *inputs, **kwargs)
        self.spatial_merge_size = config.spatial_merge_size
        self.patch_size = config.patch_size
        self.fullatt_block_indexes = config.fullatt_block_indexes
        self.window_size = config.window_size
        self.spatial_merge_unit = self.spatial_merge_size * self.spatial_merge_size

        self.patch_embed = Qwen2_5_VisionPatchEmbed(
            patch_size=config.patch_size,
            temporal_patch_size=config.temporal_patch_size,
            in_channels=config.in_channels,
            embed_dim=config.hidden_size,
        )

        head_dim = config.hidden_size // config.num_heads
        self.rotary_pos_emb = Qwen2_5_VisionRotaryEmbedding(head_dim // 2)

        self.blocks = nn.ModuleList([Qwen2_5_VLVisionBlock(config) for _ in range(config.depth)])
        self.merger = Qwen2_5_VLPatchMerger(
            dim=config.out_hidden_size,
            context_dim=config.hidden_size,
            spatial_merge_size=config.spatial_merge_size,
        )
        self.gradient_checkpointing = False

    def rot_pos_emb(self, grid_thw):
        pos_ids = []
        for t, h, w in grid_thw:
            hpos_ids = torch.arange(h).unsqueeze(1).expand(-1, w)
            hpos_ids = hpos_ids.reshape(
                h // self.spatial_merge_size,
                self.spatial_merge_size,
                w // self.spatial_merge_size,
                self.spatial_merge_size,
            )
            hpos_ids = hpos_ids.permute(0, 2, 1, 3)
            hpos_ids = hpos_ids.flatten()

            wpos_ids = torch.arange(w).unsqueeze(0).expand(h, -1)
            wpos_ids = wpos_ids.reshape(
                h // self.spatial_merge_size,
                self.spatial_merge_size,
                w // self.spatial_merge_size,
                self.spatial_merge_size,
            )
            wpos_ids = wpos_ids.permute(0, 2, 1, 3)
            wpos_ids = wpos_ids.flatten()
            pos_ids.append(torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1))
        pos_ids = torch.cat(pos_ids, dim=0)
        max_grid_size = grid_thw[:, 1:].max()
        rotary_pos_emb_full = self.rotary_pos_emb(max_grid_size)
        rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
        return rotary_pos_emb

    def get_window_index(self, grid_thw):
        window_index: list = []
        cu_window_seqlens: list = [0]
        window_index_id = 0
        vit_merger_window_size = self.window_size // self.spatial_merge_size // self.patch_size

        for grid_t, grid_h, grid_w in grid_thw:
            llm_grid_h, llm_grid_w = (
                grid_h // self.spatial_merge_size,
                grid_w // self.spatial_merge_size,
            )
            index = torch.arange(grid_t * llm_grid_h * llm_grid_w).reshape(grid_t, llm_grid_h, llm_grid_w)
            pad_h = vit_merger_window_size - llm_grid_h % vit_merger_window_size
            pad_w = vit_merger_window_size - llm_grid_w % vit_merger_window_size
            num_windows_h = (llm_grid_h + pad_h) // vit_merger_window_size
            num_windows_w = (llm_grid_w + pad_w) // vit_merger_window_size
            index_padded = F.pad(index, (0, pad_w, 0, pad_h), "constant", -100)
            index_padded = index_padded.reshape(
                grid_t,
                num_windows_h,
                vit_merger_window_size,
                num_windows_w,
                vit_merger_window_size,
            )
            index_padded = index_padded.permute(0, 1, 3, 2, 4).reshape(
                grid_t,
                num_windows_h * num_windows_w,
                vit_merger_window_size,
                vit_merger_window_size,
            )
            seqlens = (index_padded != -100).sum([2, 3]).reshape(-1)
            index_padded = index_padded.reshape(-1)
            index_new = index_padded[index_padded != -100]
            window_index.append(index_new + window_index_id)
            cu_seqlens_tmp = seqlens.cumsum(0) * self.spatial_merge_unit + cu_window_seqlens[-1]
            cu_window_seqlens.extend(cu_seqlens_tmp.tolist())
            window_index_id += (grid_t * llm_grid_h * llm_grid_w).item()
        window_index = torch.cat(window_index, dim=0)

        return window_index, cu_window_seqlens

    def forward(self, hidden_states: torch.Tensor, grid_thw: torch.Tensor, **kwargs) -> torch.Tensor:
        """
        Args:
            hidden_states (`torch.Tensor` of shape `(seq_len, hidden_size)`):
                The final hidden states of the model.
            grid_thw (`torch.Tensor` of shape `(num_images_or_videos, 3)`):
                The temporal, height and width of feature shape of each image in LLM.

        Returns:
            `torch.Tensor`: hidden_states.
        """
        hidden_states = self.patch_embed(hidden_states)
        rotary_pos_emb = self.rot_pos_emb(grid_thw)
        window_index, cu_window_seqlens = self.get_window_index(grid_thw)
        cu_window_seqlens = torch.tensor(
            cu_window_seqlens,
            device=hidden_states.device,
            dtype=grid_thw.dtype if torch.jit.is_tracing() else torch.int32,
        )
        cu_window_seqlens = torch.unique_consecutive(cu_window_seqlens)

        seq_len, _ = hidden_states.size()
        hidden_states = hidden_states.reshape(seq_len // self.spatial_merge_unit, self.spatial_merge_unit, -1)
        hidden_states = hidden_states[window_index, :, :]
        hidden_states = hidden_states.reshape(seq_len, -1)
        rotary_pos_emb = rotary_pos_emb.reshape(seq_len // self.spatial_merge_unit, self.spatial_merge_unit, -1)
        rotary_pos_emb = rotary_pos_emb[window_index, :, :]
        rotary_pos_emb = rotary_pos_emb.reshape(seq_len, -1)
        emb = torch.cat((rotary_pos_emb, rotary_pos_emb), dim=-1)
        position_embeddings = (emb.cos(), emb.sin())

        cu_seqlens = torch.repeat_interleave(grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0]).cumsum(
            dim=0,
            # Select dtype based on the following factors:
            #  - FA2 requires that cu_seqlens_q must have dtype int32
            #  - torch.onnx.export requires that cu_seqlens_q must have same dtype as grid_thw
            # See https://github.com/huggingface/transformers/pull/34852 for more information
            dtype=grid_thw.dtype if torch.jit.is_tracing() else torch.int32,
        )
        cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)

        for layer_num, blk in enumerate(self.blocks):
            if layer_num in self.fullatt_block_indexes:
                cu_seqlens_now = cu_seqlens
            else:
                cu_seqlens_now = cu_window_seqlens

            hidden_states = blk(
                hidden_states,
                cu_seqlens=cu_seqlens_now,
                position_embeddings=position_embeddings,
                **kwargs,
            )

        hidden_states = self.merger(hidden_states)
        reverse_indices = torch.argsort(window_index)
        hidden_states = hidden_states[reverse_indices, :]

        return hidden_states


@dataclass
@auto_docstring(
    custom_intro="""
    Base class for Llava outputs, with hidden states and attentions.
    """
)
class Qwen2_5_VLModelOutputWithPast(ModelOutput):
    r"""
    past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
        It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache).

        Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
        `past_key_values` input) to speed up sequential decoding.
    rope_deltas (`torch.LongTensor` of shape `(batch_size, )`, *optional*):
        The rope index difference between sequence length and multimodal rope.
    """

    last_hidden_state: Optional[torch.FloatTensor] = None
    past_key_values: Optional[Cache] = None
    hidden_states: Optional[tuple[torch.FloatTensor]] = None
    attentions: Optional[tuple[torch.FloatTensor]] = None
    rope_deltas: Optional[torch.LongTensor] = None


class Qwen2_5_VLRotaryEmbedding(nn.Module):
    inv_freq: torch.Tensor  # fix linting for `register_buffer`

    def __init__(self, config: Qwen2_5_VLTextConfig, device=None):
        super().__init__()
        # BC: "rope_type" was originally "type"
        if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
        else:
            self.rope_type = "default"
        self.max_seq_len_cached = config.max_position_embeddings
        self.original_max_seq_len = config.max_position_embeddings

        self.config = config
        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]

        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
        self.register_buffer("inv_freq", inv_freq, persistent=False)
        self.original_inv_freq = self.inv_freq

    @torch.no_grad()
    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
    def forward(self, x, position_ids):
        # In contrast to other models, Qwen2_5_VL has different position ids for the grids
        # So we expand the inv_freq to shape (3, ...)
        inv_freq_expanded = self.inv_freq[None, None, :, None].float().expand(3, position_ids.shape[1], -1, 1)
        position_ids_expanded = position_ids[:, :, None, :].float()  # shape (3, bs, 1, positions)

        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(2, 3)
            emb = torch.cat((freqs, freqs), dim=-1)
            cos = emb.cos() * self.attention_scaling
            sin = emb.sin() * self.attention_scaling

        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)


class Qwen2MLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size
        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
        self.act_fn = ACT2FN[config.hidden_act]

    def forward(self, x):
        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
        return down_proj


def apply_multimodal_rotary_pos_emb(q, k, cos, sin, mrope_section, unsqueeze_dim=1):
    """Applies Rotary Position Embedding with Multimodal Sections to the query and key tensors (https://qwenlm.github.io/blog/qwen2-vl/).

    Explanation:
        Multimodal 3D rotary position embedding is an extension to 1D rotary position embedding. The input embedding
        sequence contains vision (images / videos) embedding and text embedding or just contains text embedding. For
        vision embedding part, we apply rotary position embedding on temporal, height and width dimension separately.
        Here we split the channel dimension to 3 chunks for the temporal, height and width rotary position embedding.
        For text embedding part, we just apply 1D rotary position embedding. The three rotary position index (temporal,
        height and width) of text embedding is always the same, so the text embedding rotary position embedding has no
        difference with modern LLMs.

    Args:
        q (`torch.Tensor`): The query tensor.
        k (`torch.Tensor`): The key tensor.
        cos (`torch.Tensor`): The cosine part of the rotary embedding.
        sin (`torch.Tensor`): The sine part of the rotary embedding.
        position_ids (`torch.Tensor`):
            The position indices of the tokens corresponding to the query and key tensors. For example, this can be
            used to pass offsetted position ids when working with a KV-cache.
        mrope_section(`List(int)`):
            Multimodal rope section is for channel dimension of temporal, height and width in rope calculation.
        unsqueeze_dim (`int`, *optional*, defaults to 1):
            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
    Returns:
        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
    """
    mrope_section = mrope_section * 2
    cos = torch.cat([m[i % 3] for i, m in enumerate(cos.split(mrope_section, dim=-1))], dim=-1).unsqueeze(
        unsqueeze_dim
    )
    sin = torch.cat([m[i % 3] for i, m in enumerate(sin.split(mrope_section, dim=-1))], dim=-1).unsqueeze(
        unsqueeze_dim
    )

    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    return q_embed, k_embed


class Qwen2_5_VLAttention(nn.Module):
    """
    Multi-headed attention from 'Attention Is All You Need' paper. Modified to use sliding window attention: Longformer
    and "Generating Long Sequences with Sparse Transformers".
    """

    def __init__(self, config: Qwen2_5_VLTextConfig, layer_idx: Optional[int] = None):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        if layer_idx is None:
            logger.warning_once(
                f"Instantiating {self.__class__.__name__} without passing `layer_idx` is not recommended and will "
                "to errors during the forward call, if caching is used. Please make sure to provide a `layer_idx` "
                "when creating this class."
            )

        self.hidden_size = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.head_dim = self.hidden_size // self.num_heads
        self.num_key_value_heads = config.num_key_value_heads
        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
        self.is_causal = True
        self.attention_dropout = config.attention_dropout
        self.rope_scaling = config.rope_scaling
        self.scaling = self.head_dim**-0.5

        if (self.head_dim * self.num_heads) != self.hidden_size:
            raise ValueError(
                f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
                f" and `num_heads`: {self.num_heads})."
            )
        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=True)
        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True)
        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True)
        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
        self.sliding_window = config.sliding_window if config.layer_types[layer_idx] == "sliding_attention" else None

        self.rotary_emb = Qwen2_5_VLRotaryEmbedding(config=config)

    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Cache] = None,
        output_attentions: bool = False,
        use_cache: bool = False,
        cache_position: Optional[torch.LongTensor] = None,
        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]:
        bsz, q_len, _ = hidden_states.size()

        query_states = self.q_proj(hidden_states)
        key_states = self.k_proj(hidden_states)
        value_states = self.v_proj(hidden_states)

        query_states = query_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)
        key_states = key_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)
        value_states = value_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)

        cos, sin = position_embeddings
        query_states, key_states = apply_multimodal_rotary_pos_emb(
            query_states, key_states, cos, sin, self.rope_scaling["mrope_section"]
        )

        if past_key_values is not None:
            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}  # Specific to RoPE models
            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)

        attention_interface: Callable = eager_attention_forward
        if self.config._attn_implementation != "eager":
            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]

        attn_output, attn_weights = attention_interface(
            self,
            query_states,
            key_states,
            value_states,
            attention_mask,
            dropout=0.0 if not self.training else self.attention_dropout,
            scaling=self.scaling,
            sliding_window=self.sliding_window,
            position_ids=position_ids,  # pass positions for FA2
            **kwargs,
        )

        attn_output = attn_output.reshape(bsz, q_len, -1).contiguous()
        attn_output = self.o_proj(attn_output)
        return attn_output, attn_weights


class Qwen2_5_VLDecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: Qwen2_5_VLTextConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size

        if config.use_sliding_window and config._attn_implementation != "flash_attention_2":
            logger.warning_once(
                f"Sliding Window Attention is enabled but not implemented for `{config._attn_implementation}`; "
                "unexpected results may be encountered."
            )
        self.self_attn = Qwen2_5_VLAttention(config, layer_idx)

        self.mlp = Qwen2MLP(config)
        self.input_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.attention_type = config.layer_types[layer_idx]

    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Cache] = None,
        output_attentions: Optional[bool] = False,
        use_cache: Optional[bool] = False,
        cache_position: Optional[torch.LongTensor] = None,
        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> tuple[torch.FloatTensor, Optional[tuple[torch.FloatTensor, torch.FloatTensor]]]:
        """
        Args:
            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
                `(batch, sequence_length)` where padding elements are indicated by 0.
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
            use_cache (`bool`, *optional*):
                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
                (see `past_key_values`).
            past_key_values (`Cache`, *optional*): cached past key and value projection states
            cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
                Indices depicting the position of the input sequence tokens in the sequence.
            position_embeddings (`tuple[torch.FloatTensor, torch.FloatTensor]`, *optional*):
                Tuple containing the cosine and sine positional embeddings of shape `(batch_size, seq_len, head_dim)`,
                with `head_dim` being the embedding dimension of each attention head.
            kwargs (`dict`, *optional*):
                Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code
                into the model
        """

        residual = hidden_states

        hidden_states = self.input_layernorm(hidden_states)

        # Self Attention
        hidden_states, self_attn_weights = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            output_attentions=output_attentions,
            use_cache=use_cache,
            cache_position=cache_position,
            position_embeddings=position_embeddings,
            **kwargs,
        )
        hidden_states = residual + hidden_states

        # Fully Connected
        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (self_attn_weights,)

        return outputs


@auto_docstring
class Qwen2_5_VLTextModel(Qwen2_5_VLPreTrainedModel):
    config: Qwen2_5_VLTextConfig

    def __init__(self, config: Qwen2_5_VLTextConfig):
        super().__init__(config)
        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size

        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
        self.layers = nn.ModuleList(
            [Qwen2_5_VLDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self._attn_implementation = config._attn_implementation
        self.norm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = Qwen2_5_VLRotaryEmbedding(config=config)
        self.has_sliding_layers = "sliding_attention" in self.config.layer_types

        self.gradient_checkpointing = False
        # Initialize weights and apply final processing
        self.post_init()

    @auto_docstring
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Cache] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> Union[tuple, BaseModelOutputWithPast]:
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache

        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

        if self.gradient_checkpointing and self.training:
            if use_cache:
                logger.warning_once(
                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                )
                use_cache = False

        # torch.jit.trace() doesn't support cache objects in the output
        if use_cache and past_key_values is None and not torch.jit.is_tracing():
            past_key_values = DynamicCache(config=self.config)

        if inputs_embeds is None:
            inputs_embeds = self.embed_tokens(input_ids)

        if cache_position is None:
            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
            cache_position = torch.arange(
                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
            )

        # the hard coded `3` is for temporal, height and width.
        if position_ids is None:
            position_ids = cache_position.view(1, 1, -1).expand(3, inputs_embeds.shape[0], -1)
        elif position_ids.ndim == 2:
            position_ids = position_ids[None, ...].expand(3, position_ids.shape[0], -1)

        # NOTE: we need to pass text position ids for packing. Qwen2-VL uses 3D positions
        # where each dim indicates visual spatial positions for temporal/height/width grids.
        # There are two scenarios when FA2-like packed masking might be activated.
        # 1. User specifically passed packed `position_ids` and no attention mask.
        #    In this case we expect the useer to create correct position ids for all 3 grids
        #    and prepend text-only position ids to it. The final tensor will be [4, bs, seq-len]
        # 2. User runs forward with no attention mask and no position ids. In this case, position ids
        #    are prepared by the model (`get_rope_index`) as `[4, bs, seq-len]` tensor. Text-only positions are
        #    prepended by us when creating positions so that the mask is constructed correctly. NOTE: failing to pass
        #    text-only positions will cause incorrect mask construction, do not change `prepare_input_for_generation`
        if position_ids.ndim == 3 and position_ids.shape[0] == 4:
            text_position_ids = position_ids[0]
            position_ids = position_ids[1:]
        else:
            # If inputs are not packed (usual 3D positions), do not prepare mask from position_ids
            text_position_ids = None

        # It may already have been prepared by e.g. `generate`
        if not isinstance(causal_mask_mapping := attention_mask, dict):
            # Prepare mask arguments
            mask_kwargs = {
                "config": self.config,
                "input_embeds": inputs_embeds,
                "attention_mask": attention_mask,
                "cache_position": cache_position,
                "past_key_values": past_key_values,
                "position_ids": text_position_ids,
            }
            # Create the masks
            causal_mask_mapping = {
                "full_attention": create_causal_mask(**mask_kwargs),
            }
            # The sliding window alternating layers are not always activated depending on the config
            if self.has_sliding_layers:
                causal_mask_mapping["sliding_attention"] = create_sliding_window_causal_mask(**mask_kwargs)

        hidden_states = inputs_embeds

        # create position embeddings to be shared across the decoder layers
        position_embeddings = self.rotary_emb(hidden_states, position_ids)

        # decoder layers
        all_hidden_states = () if output_hidden_states else None
        all_self_attns = () if output_attentions else None

        for decoder_layer in self.layers:
            if output_hidden_states:
                all_hidden_states += (hidden_states,)

            layer_outputs = decoder_layer(
                hidden_states,
                attention_mask=causal_mask_mapping[decoder_layer.attention_type],
                position_ids=text_position_ids,
                past_key_values=past_key_values,
                output_attentions=output_attentions,
                use_cache=use_cache,
                cache_position=cache_position,
                position_embeddings=position_embeddings,
                **kwargs,
            )

            hidden_states = layer_outputs[0]

            if output_attentions:
                all_self_attns += (layer_outputs[1],)

        hidden_states = self.norm(hidden_states)

        # add hidden states from the last decoder layer
        if output_hidden_states:
            all_hidden_states += (hidden_states,)

        if not return_dict:
            return tuple(
                v for v in [hidden_states, past_key_values, all_hidden_states, all_self_attns] if v is not None
            )
        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values,
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
        )


@auto_docstring
class Qwen2_5_VLModel(Qwen2_5_VLPreTrainedModel):
    base_model_prefix = ""
    _checkpoint_conversion_mapping = {"^model": "language_model"}
    # Reference: fix gemma3 grad acc #37208
    accepts_loss_kwargs = False
    config: Qwen2_5_VLConfig
    _no_split_modules = ["Qwen2_5_VLDecoderLayer", "Qwen2_5_VLVisionBlock"]

    def __init__(self, config):
        super().__init__(config)
        self.visual = Qwen2_5_VisionTransformerPretrainedModel._from_config(config.vision_config)
        self.language_model = Qwen2_5_VLTextModel._from_config(config.text_config)
        self.rope_deltas = None  # cache rope_deltas here
        # Initialize weights and apply final processing
        self.post_init()

    def get_input_embeddings(self):
        return self.language_model.get_input_embeddings()

    def set_input_embeddings(self, value):
        self.language_model.set_input_embeddings(value)

    def set_decoder(self, decoder):
        self.language_model = decoder

    def get_decoder(self):
        return self.language_model

    def get_rope_index(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        image_grid_thw: Optional[torch.LongTensor] = None,
        video_grid_thw: Optional[torch.LongTensor] = None,
        second_per_grid_ts: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        """
        Calculate the 3D rope index based on image and video's temporal, height and width in LLM.

        Explanation:
            Each embedding sequence contains vision embedding and text embedding or just contains text embedding.

            For pure text embedding sequence, the rotary position embedding has no difference with modern LLMs.
            Examples:
                input_ids: [T T T T T], here T is for text.
                temporal position_ids: [0, 1, 2, 3, 4]
                height position_ids: [0, 1, 2, 3, 4]
                width position_ids: [0, 1, 2, 3, 4]

            For vision and text embedding sequence, we calculate 3D rotary position embedding for vision part
            and 1D rotary position embedding for text part.
            Examples:
                Temporal (Time): 3 patches, representing different segments of the video in time.
                Height: 2 patches, dividing each frame vertically.
                Width: 2 patches, dividing each frame horizontally.
                We also have some important parameters:
                fps (Frames Per Second): The video's frame rate, set to 1. This means one frame is processed each second.
                tokens_per_second: This is a crucial parameter. It dictates how many "time-steps" or "temporal tokens" are conceptually packed into a one-second interval of the video. In this case, we have 25 tokens per second. So each second of the video will be represented with 25 separate time points. It essentially defines the temporal granularity.
                temporal_patch_size: The number of frames that compose one temporal patch. Here, it's 2 frames.
                interval: The step size for the temporal position IDs, calculated as tokens_per_second * temporal_patch_size / fps. In this case, 25 * 2 / 1 = 50. This means that each temporal patch will be have a difference of 50 in the temporal position IDs.
                input_ids: [V V V V V V V V V V V V T T T T T], here V is for vision.
                vision temporal position_ids: [0, 0, 0, 0, 50, 50, 50, 50, 100, 100, 100, 100]
                vision height position_ids: [0, 0, 1, 1, 0, 0, 1, 1, 0, 0, 1, 1]
                vision width position_ids: [0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1]
                text temporal position_ids: [101, 102, 103, 104, 105]
                text height position_ids: [101, 102, 103, 104, 105]
                text width position_ids: [101, 102, 103, 104, 105]
                Here we calculate the text start position_ids as the max vision position_ids plus 1.

        Args:
            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
                it.
            image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
                The temporal, height and width of feature shape of each image in LLM.
            video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
                The temporal, height and width of feature shape of each video in LLM.
            second_per_grid_ts (`torch.Tensor` of shape `(num_videos)`, *optional*):
                The time interval (in seconds) for each grid along the temporal dimension in the 3D position IDs.
            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:

                - 1 for tokens that are **not masked**,
                - 0 for tokens that are **masked**.

        Returns:
            position_ids (`torch.LongTensor` of shape `(3, batch_size, sequence_length)`)
            mrope_position_deltas (`torch.Tensor` of shape `(batch_size)`)
        """
        spatial_merge_size = self.config.vision_config.spatial_merge_size
        image_token_id = self.config.image_token_id
        video_token_id = self.config.video_token_id
        vision_start_token_id = self.config.vision_start_token_id
        mrope_position_deltas = []
        if input_ids is not None and (image_grid_thw is not None or video_grid_thw is not None):
            total_input_ids = input_ids
            if attention_mask is not None:
                attention_mask = attention_mask == 1
            position_ids = torch.ones(
                3,
                input_ids.shape[0],
                input_ids.shape[1],
                dtype=input_ids.dtype,
                device=input_ids.device,
            )
            image_index, video_index = 0, 0
            for i, input_ids in enumerate(total_input_ids):
                if attention_mask is not None:
                    input_ids = input_ids[attention_mask[i]]
                image_nums, video_nums = 0, 0
                vision_start_indices = torch.argwhere(input_ids == vision_start_token_id).squeeze(1)
                vision_tokens = input_ids[vision_start_indices + 1]
                image_nums = (vision_tokens == image_token_id).sum()
                video_nums = (vision_tokens == video_token_id).sum()
                input_tokens = input_ids.tolist()
                llm_pos_ids_list: list = []
                st = 0
                remain_images, remain_videos = image_nums, video_nums
                for _ in range(image_nums + video_nums):
                    if image_token_id in input_tokens and remain_images > 0:
                        ed_image = input_tokens.index(image_token_id, st)
                    else:
                        ed_image = len(input_tokens) + 1
                    if video_token_id in input_tokens and remain_videos > 0:
                        ed_video = input_tokens.index(video_token_id, st)
                    else:
                        ed_video = len(input_tokens) + 1
                    if ed_image < ed_video:
                        t, h, w = (
                            image_grid_thw[image_index][0],
                            image_grid_thw[image_index][1],
                            image_grid_thw[image_index][2],
                        )
                        second_per_grid_t = 0
                        image_index += 1
                        remain_images -= 1
                        ed = ed_image

                    else:
                        t, h, w = (
                            video_grid_thw[video_index][0],
                            video_grid_thw[video_index][1],
                            video_grid_thw[video_index][2],
                        )
                        if second_per_grid_ts is not None:
                            second_per_grid_t = second_per_grid_ts[video_index]
                        else:
                            second_per_grid_t = 1.0
                        video_index += 1
                        remain_videos -= 1
                        ed = ed_video
                    llm_grid_t, llm_grid_h, llm_grid_w = (
                        t.item(),
                        h.item() // spatial_merge_size,
                        w.item() // spatial_merge_size,
                    )
                    text_len = ed - st

                    st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
                    llm_pos_ids_list.append(torch.arange(text_len).view(1, -1).expand(3, -1) + st_idx)

                    range_tensor = torch.arange(llm_grid_t).view(-1, 1)
                    expanded_range = range_tensor.expand(-1, llm_grid_h * llm_grid_w)

                    ## normalize type, send to device.
                    second_per_grid_t = torch.as_tensor(
                        second_per_grid_t, dtype=range_tensor.dtype, device=range_tensor.device
                    )

                    time_tensor = expanded_range * second_per_grid_t * self.config.vision_config.tokens_per_second

                    time_tensor_long = time_tensor.long()
                    t_index = time_tensor_long.flatten()

                    h_index = torch.arange(llm_grid_h).view(1, -1, 1).expand(llm_grid_t, -1, llm_grid_w).flatten()
                    w_index = torch.arange(llm_grid_w).view(1, 1, -1).expand(llm_grid_t, llm_grid_h, -1).flatten()
                    llm_pos_ids_list.append(torch.stack([t_index, h_index, w_index]) + text_len + st_idx)
                    st = ed + llm_grid_t * llm_grid_h * llm_grid_w

                if st < len(input_tokens):
                    st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
                    text_len = len(input_tokens) - st
                    llm_pos_ids_list.append(torch.arange(text_len).view(1, -1).expand(3, -1) + st_idx)

                llm_positions = torch.cat(llm_pos_ids_list, dim=1).reshape(3, -1)
                if attention_mask is not None:
                    position_ids[..., i, attention_mask[i]] = llm_positions.to(position_ids.device)
                else:
                    position_ids[..., i, :] = llm_positions.to(position_ids.device)
                mrope_position_deltas.append(llm_positions.max() + 1 - len(total_input_ids[i]))
            mrope_position_deltas = torch.tensor(mrope_position_deltas).unsqueeze(1).to(device=input_ids.device)
            return position_ids, mrope_position_deltas
        else:
            if attention_mask is not None:
                position_ids = attention_mask.long().cumsum(-1) - 1
                position_ids.masked_fill_(attention_mask == 0, 1)
                position_ids = position_ids.unsqueeze(0).expand(3, -1, -1).to(attention_mask.device)
                max_position_ids = position_ids.max(0, keepdim=False)[0].max(-1, keepdim=True)[0]
                mrope_position_deltas = max_position_ids + 1 - attention_mask.shape[-1]
            else:
                position_ids = (
                    torch.arange(input_ids.shape[1], device=input_ids.device)
                    .view(1, 1, -1)
                    .expand(3, input_ids.shape[0], -1)
                )
                mrope_position_deltas = torch.zeros(
                    [input_ids.shape[0], 1],
                    device=input_ids.device,
                    dtype=input_ids.dtype,
                )

            return position_ids, mrope_position_deltas

    def get_video_features(
        self, pixel_values_videos: torch.FloatTensor, video_grid_thw: Optional[torch.LongTensor] = None
    ):
        """
        Encodes videos into continuous embeddings that can be forwarded to the language model.

        Args:
            pixel_values_videos (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
                The tensors corresponding to the input videos.
            video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
                The temporal, height and width of feature shape of each video in LLM.
        """
        pixel_values_videos = pixel_values_videos.type(self.visual.dtype)
        video_embeds = self.visual(pixel_values_videos, grid_thw=video_grid_thw)
        split_sizes = (video_grid_thw.prod(-1) // self.visual.spatial_merge_size**2).tolist()
        video_embeds = torch.split(video_embeds, split_sizes)
        return video_embeds

    def get_image_features(self, pixel_values: torch.FloatTensor, image_grid_thw: Optional[torch.LongTensor] = None):
        """
        Encodes images into continuous embeddings that can be forwarded to the language model.

        Args:
            pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
                The tensors corresponding to the input images.
            image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
                The temporal, height and width of feature shape of each image in LLM.
        """
        pixel_values = pixel_values.type(self.visual.dtype)
        image_embeds = self.visual(pixel_values, grid_thw=image_grid_thw)
        split_sizes = (image_grid_thw.prod(-1) // self.visual.spatial_merge_size**2).tolist()
        image_embeds = torch.split(image_embeds, split_sizes)
        return image_embeds

    def get_placeholder_mask(
        self,
        input_ids: torch.LongTensor,
        inputs_embeds: torch.FloatTensor,
        image_features: Optional[torch.FloatTensor] = None,
        video_features: Optional[torch.FloatTensor] = None,
    ):
        """
        Obtains multimodal placeholder mask from `input_ids` or `inputs_embeds`, and checks that the placeholder token count is
        equal to the length of multimodal features. If the lengths are different, an error is raised.
        """
        if input_ids is None:
            special_image_mask = inputs_embeds == self.get_input_embeddings()(
                torch.tensor(self.config.image_token_id, dtype=torch.long, device=inputs_embeds.device)
            )
            special_image_mask = special_image_mask.all(-1)
            special_video_mask = inputs_embeds == self.get_input_embeddings()(
                torch.tensor(self.config.video_token_id, dtype=torch.long, device=inputs_embeds.device)
            )
            special_video_mask = special_video_mask.all(-1)
        else:
            special_image_mask = input_ids == self.config.image_token_id
            special_video_mask = input_ids == self.config.video_token_id

        n_image_tokens = special_image_mask.sum()
        special_image_mask = special_image_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
        if image_features is not None and inputs_embeds[special_image_mask].numel() != image_features.numel():
            raise ValueError(
                f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {image_features.shape[0]}"
            )

        n_video_tokens = special_video_mask.sum()
        special_video_mask = special_video_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
        if video_features is not None and inputs_embeds[special_video_mask].numel() != video_features.numel():
            raise ValueError(
                f"Videos features and video tokens do not match: tokens: {n_video_tokens}, features {video_features.shape[0]}"
            )

        return special_image_mask, special_video_mask

    def get_query_placeholder_mask(
        self,
        input_ids: torch.LongTensor,
        inputs_embeds: torch.FloatTensor,
    ):
        """
        Obtains multimodal placeholder mask from `input_ids` or `inputs_embeds`, and checks that the placeholder token count is
        equal to the length of multimodal features. If the lengths are different, an error is raised.
        """
        if input_ids is None:
            special_query_mask = inputs_embeds == self.get_input_embeddings()(
                torch.tensor(151646, dtype=torch.long, device=inputs_embeds.device)
            )
            special_query_mask = special_query_mask.all(-1)
        else:
            special_query_mask = input_ids == 151646
        n_query_tokens = special_query_mask.sum() // inputs_embeds.size(0)
        special_query_mask = special_query_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
        # print(special_query_mask.size())
        if (n_query_tokens == self.num_image_queries) or (n_query_tokens == self.num_video_queries)  or (n_query_tokens == (self.num_video_queries + self.num_ref_queries)) or (n_query_tokens == (self.num_ref_queries)):
            return special_query_mask
        else:
            raise ValueError(
                f"Learnable Query and image tokens do not match: tokens: {n_query_tokens}"
            )

    @auto_docstring
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Cache] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        pixel_values: Optional[torch.Tensor] = None,
        pixel_values_videos: Optional[torch.FloatTensor] = None,
        image_grid_thw: Optional[torch.LongTensor] = None,
        video_grid_thw: Optional[torch.LongTensor] = None,
        rope_deltas: Optional[torch.LongTensor] = None,
        cache_position: Optional[torch.LongTensor] = None,
        second_per_grid_ts: Optional[torch.Tensor] = None,
        learnable_query: Optional[torch.FloatTensor] = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> Union[tuple, Qwen2_5_VLModelOutputWithPast]:
        r"""
        image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
            The temporal, height and width of feature shape of each image in LLM.
        video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
            The temporal, height and width of feature shape of each video in LLM.
        rope_deltas (`torch.LongTensor` of shape `(batch_size, )`, *optional*):
            The rope index difference between sequence length and multimodal rope.
        second_per_grid_ts (`torch.Tensor` of shape `(num_videos)`, *optional*):
            The time interval (in seconds) for each grid along the temporal dimension in the 3D position IDs.
        """

        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        if inputs_embeds is None:
            inputs_embeds = self.get_input_embeddings()(input_ids)

        if pixel_values is not None:
            image_embeds = self.get_image_features(pixel_values, image_grid_thw)
            image_embeds = torch.cat(image_embeds, dim=0).to(inputs_embeds.device, inputs_embeds.dtype)
            image_mask, _ = self.get_placeholder_mask(
                input_ids, inputs_embeds=inputs_embeds, image_features=image_embeds
            )
            inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_embeds)

        if pixel_values_videos is not None:
            video_embeds = self.get_video_features(pixel_values_videos, video_grid_thw)
            video_embeds = torch.cat(video_embeds, dim=0).to(inputs_embeds.device, inputs_embeds.dtype)
            _, video_mask = self.get_placeholder_mask(
                input_ids, inputs_embeds=inputs_embeds, video_features=video_embeds
            )
            inputs_embeds = inputs_embeds.masked_scatter(video_mask, video_embeds)
        # print(inputs_embeds.shape, attention_mask.shape, learnable_query.shape)
        if learnable_query is not None:
            query_mask = self.get_query_placeholder_mask(input_ids, inputs_embeds=inputs_embeds)
            inputs_embeds = inputs_embeds.masked_scatter(query_mask, learnable_query.unsqueeze(0))

        if position_ids is None:
            # Calculate RoPE index once per generation in the pre-fill stage only.
            # When compiling, we can't check tensor values thus we check only input length
            # It is safe to assume that `length!=1` means we're in pre-fill because compiled
            # models currently cannot do asssisted decoding
            prefill_compiled_stage = is_torchdynamo_compiling() and (
                (input_ids is not None and input_ids.shape[1] != 1)
                or (inputs_embeds is not None and inputs_embeds.shape[1] != 1)
            )
            prefill_noncompiled_stage = not is_torchdynamo_compiling() and (
                (cache_position is not None and cache_position[0] == 0)
                or (past_key_values is None or past_key_values.get_seq_length() == 0)
            )
            if (prefill_compiled_stage or prefill_noncompiled_stage) or self.rope_deltas is None:
                position_ids, rope_deltas = self.get_rope_index(
                    input_ids,
                    image_grid_thw,
                    video_grid_thw,
                    second_per_grid_ts=second_per_grid_ts,
                    attention_mask=attention_mask,
                )
                self.rope_deltas = rope_deltas
            else:
                batch_size, seq_length, _ = inputs_embeds.shape
                position_ids = torch.arange(seq_length, device=inputs_embeds.device)
                position_ids = position_ids.view(1, 1, -1).expand(3, batch_size, -1)
                if cache_position is not None:
                    delta = (cache_position[0] + self.rope_deltas).to(inputs_embeds.device)
                else:
                    delta = torch.zeros((batch_size, seq_length), device=inputs_embeds.device)
                delta = delta.repeat_interleave(batch_size // delta.shape[0], dim=1)
                position_ids = position_ids + delta.to(position_ids.device)

        outputs = self.language_model(
            input_ids=None,
            position_ids=position_ids,
            attention_mask=attention_mask,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=True,
            cache_position=cache_position,
            **kwargs,
        )

        output = Qwen2_5_VLModelOutputWithPast(
            last_hidden_state=outputs.last_hidden_state,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            rope_deltas=self.rope_deltas,
        )
        return output if return_dict else (output.last_hidden_state, output.past_key_values, output.hidden_states, output.attentions, output.rope_deltas)


@dataclass
@auto_docstring(
    custom_intro="""
    Base class for Qwen2_5_VL causal language model (or autoregressive) outputs.
    """
)
class Qwen2_5_VLCausalLMOutputWithPast(ModelOutput):
    r"""
    loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
        Language modeling loss (for next-token prediction).
    logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
        Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
    past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
        It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache).

        Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
        `past_key_values` input) to speed up sequential decoding.
    rope_deltas (`torch.LongTensor` of shape `(batch_size, )`, *optional*):
        The rope index difference between sequence length and multimodal rope.
    """

    loss: Optional[torch.FloatTensor] = None
    logits: Optional[torch.FloatTensor] = None
    past_key_values: Optional[Cache] = None
    hidden_states: Optional[tuple[torch.FloatTensor]] = None
    attentions: Optional[tuple[torch.FloatTensor]] = None
    rope_deltas: Optional[torch.LongTensor] = None


class Qwen2_5_VLForConditionalGeneration(Qwen2_5_VLPreTrainedModel, GenerationMixin):
    _checkpoint_conversion_mapping = {
        "^visual": "model.visual",
        r"^model(?!\.(language_model|visual))": "model.language_model",
    }
    _tied_weights_keys = ["lm_head.weight"]
    # Reference: fix gemma3 grad acc #37208
    accepts_loss_kwargs = False

    def __init__(self, config):
        super().__init__(config)
        self.model = Qwen2_5_VLModel(config)
        # self.lm_head = nn.Linear(config.text_config.hidden_size, config.text_config.vocab_size, bias=False)

        self.post_init()

    def get_input_embeddings(self):
        return self.model.get_input_embeddings()

    def set_input_embeddings(self, value):
        self.model.set_input_embeddings(value)

    def set_decoder(self, decoder):
        self.model.set_decoder(decoder)

    def get_decoder(self):
        return self.model.get_decoder()

    def get_video_features(
        self, pixel_values_videos: torch.FloatTensor, video_grid_thw: Optional[torch.LongTensor] = None
    ):
        return self.model.get_video_features(pixel_values_videos, video_grid_thw)

    def get_image_features(self, pixel_values: torch.FloatTensor, image_grid_thw: Optional[torch.LongTensor] = None):
        return self.model.get_image_features(pixel_values, image_grid_thw)

    # Make modules available through conditional class for BC
    @property
    def language_model(self):
        return self.model.language_model

    @property
    def visual(self):
        return self.model.visual

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Cache] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        pixel_values: Optional[torch.Tensor] = None,
        pixel_values_videos: Optional[torch.FloatTensor] = None,
        image_grid_thw: Optional[torch.LongTensor] = None,
        video_grid_thw: Optional[torch.LongTensor] = None,
        rope_deltas: Optional[torch.LongTensor] = None,
        cache_position: Optional[torch.LongTensor] = None,
        second_per_grid_ts: Optional[torch.Tensor] = None,
        logits_to_keep: Union[int, torch.Tensor] = 0,
        **kwargs: Unpack[TransformersKwargs],
    ) -> Union[tuple, Qwen2_5_VLCausalLMOutputWithPast]:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
        image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
            The temporal, height and width of feature shape of each image in LLM.
        video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
            The temporal, height and width of feature shape of each video in LLM.
        rope_deltas (`torch.LongTensor` of shape `(batch_size, )`, *optional*):
            The rope index difference between sequence length and multimodal rope.
        second_per_grid_ts (`torch.Tensor` of shape `(num_videos)`, *optional*):
            The time interval (in seconds) for each grid along the temporal dimension in the 3D position IDs.

        Example:

        ```python
        >>> from PIL import Image
        >>> import requests
        >>> from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration

        >>> model = Qwen2_5_VLForConditionalGeneration.from_pretrained("Qwen/Qwen2.5-VL-7B-Instruct")
        >>> processor = AutoProcessor.from_pretrained("Qwen/Qwen2.5-VL-7B-Instruct")

        >>> messages = [
            {
                "role": "user",
                "content": [
                    {"type": "image"},
                    {"type": "text", "text": "What is shown in this image?"},
                ],
            },
        ]
        >>> url = "https://www.ilankelman.org/stopsigns/australia.jpg"
        >>> image = Image.open(requests.get(url, stream=True).raw)

        >>> text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
        >>> inputs = processor(text=[text], images=[image], vision_infos=[vision_infos])

        >>> # Generate
        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
        "The image shows a street scene with a red stop sign in the foreground. In the background, there is a large red gate with Chinese characters ..."
        ```"""

        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )

        outputs = self.model(
            input_ids=input_ids,
            pixel_values=pixel_values,
            pixel_values_videos=pixel_values_videos,
            image_grid_thw=image_grid_thw,
            video_grid_thw=video_grid_thw,
            second_per_grid_ts=second_per_grid_ts,
            position_ids=position_ids,
            attention_mask=attention_mask,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=True,
            cache_position=cache_position,
            **kwargs,
        )

        # hidden_states = outputs[0]
        logits = None
        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
        # slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
        # logits = self.lm_head(hidden_states[:, slice_indices, :])

        loss = None
        # if labels is not None:
        #     loss = self.loss_function(
        #         logits=logits, labels=labels, vocab_size=self.config.text_config.vocab_size, **kwargs
        #     )

        return Qwen2_5_VLCausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            rope_deltas=outputs.rope_deltas,
        )

    def prepare_inputs_for_generation(
        self,
        input_ids,
        past_key_values=None,
        attention_mask=None,
        inputs_embeds=None,
        cache_position=None,
        position_ids=None,
        use_cache=True,
        pixel_values=None,
        pixel_values_videos=None,
        image_grid_thw=None,
        video_grid_thw=None,
        second_per_grid_ts=None,
        **kwargs,
    ):
        # Overwritten -- in specific circumstances we don't want to forward image inputs to the model

        model_inputs = super().prepare_inputs_for_generation(
            input_ids,
            past_key_values=past_key_values,
            attention_mask=attention_mask,
            inputs_embeds=inputs_embeds,
            cache_position=cache_position,
            position_ids=position_ids,
            pixel_values=pixel_values,
            pixel_values_videos=pixel_values_videos,
            image_grid_thw=image_grid_thw,
            video_grid_thw=video_grid_thw,
            second_per_grid_ts=second_per_grid_ts,
            use_cache=use_cache,
            **kwargs,
        )

        # Qwen2-5-VL position_ids are prepared with rope_deltas
        if position_ids is None:
            # Calculate RoPE index once per generation in the pre-fill stage only.
            # When compiling, we can't check tensor values thus we check only input length
            # It is safe to assume that `length!=1` means we're in pre-fill because compiled
            # models currently cannot do assisted decoding
            if cache_position[0] == 0 or self.model.rope_deltas is None:
                vision_positions, rope_deltas = self.model.get_rope_index(
                    model_inputs.get("input_ids", None),
                    image_grid_thw=image_grid_thw,
                    video_grid_thw=video_grid_thw,
                    second_per_grid_ts=second_per_grid_ts,
                    attention_mask=attention_mask,
                )
                self.model.rope_deltas = rope_deltas
            # then use the prev pre-calculated rope-deltas to get the correct position ids
            elif "position_ids" in model_inputs:
                batch_size, seq_length = model_inputs["position_ids"].shape
                device = model_inputs["position_ids"].device
                position_ids = torch.arange(seq_length, device=device)
                position_ids = position_ids.view(1, 1, -1).expand(3, batch_size, -1)
                delta = cache_position[0] + self.model.rope_deltas
                delta = delta.repeat_interleave(batch_size // delta.shape[0], dim=0)
                vision_positions = position_ids + delta.expand_as(position_ids)

            # Concatenate "text + vision" positions into [4, bs, seq-len]
            text_positions = model_inputs["position_ids"][None, ...]
            model_inputs["position_ids"] = torch.cat([text_positions, vision_positions], dim=0)

        if cache_position[0] != 0:
            model_inputs["pixel_values"] = None
            model_inputs["pixel_values_videos"] = None

        return model_inputs

    def _get_image_nums_and_video_nums(
        self,
        input_ids: Optional[torch.LongTensor],
        inputs_embeds: Optional[torch.Tensor] = None,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        """
        Get the number of images and videos for each sample to calculate the separation length of the sample tensor.
        These parameters are not passed through the processor to avoid unpredictable impacts from interface modifications.

        Args:
            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                Indices of input sequence tokens in the vocabulary.

        Returns:
            image_nums (`torch.LongTensor` of shape `(batch_size, num_images_sample)`)
            video_nums (`torch.LongTensor` of shape `(batch_size, num_videos_sample)`)
        """
        image_token_id = self.config.image_token_id
        video_token_id = self.config.video_token_id
        vision_start_token_id = self.config.vision_start_token_id

        if inputs_embeds is not None:
            vision_start_mask = (
                inputs_embeds
                == self.get_input_embeddings()(
                    torch.tensor(vision_start_token_id, dtype=torch.long, device=inputs_embeds.device)
                )
            )[..., 0]
            image_mask = (
                inputs_embeds
                == self.get_input_embeddings()(
                    torch.tensor(image_token_id, dtype=torch.long, device=inputs_embeds.device)
                )
            )[..., 0]
            video_mask = (
                inputs_embeds
                == self.get_input_embeddings()(
                    torch.tensor(video_token_id, dtype=torch.long, device=inputs_embeds.device)
                )
            )[..., 0]
        else:
            vision_start_mask = input_ids == vision_start_token_id
            image_mask = input_ids == image_token_id
            video_mask = input_ids == video_token_id

        vision_first_mask = torch.roll(vision_start_mask, shifts=1, dims=1)
        image_nums = torch.sum(vision_first_mask & image_mask, dim=1)
        video_nums = torch.sum(vision_first_mask & video_mask, dim=1)

        return image_nums, video_nums

    def _expand_inputs_for_generation(
        self,
        expand_size: int = 1,
        is_encoder_decoder: bool = False,
        input_ids: Optional[torch.LongTensor] = None,
        **model_kwargs,
    ) -> tuple[torch.LongTensor, dict[str, Any]]:
        # Overwritten -- Support for expanding tensors without a batch size dimension
        # e.g., pixel_values, image_grid_thw, pixel_values_videos, video_grid_thw, second_per_grid_t
        # pixel_values.shape[0] is sum(seqlen_images for samples)
        # image_grid_thw.shape[0] is sum(num_images for samples)

        if expand_size == 1:
            return input_ids, model_kwargs

        visual_keys = ["pixel_values", "image_grid_thw", "pixel_values_videos", "video_grid_thw", "second_per_grid_ts"]

        def _expand_dict_for_generation_visual(dict_to_expand):
            image_grid_thw = model_kwargs.get("image_grid_thw", None)
            video_grid_thw = model_kwargs.get("video_grid_thw", None)
            image_nums, video_nums = self._get_image_nums_and_video_nums(
                input_ids, inputs_embeds=model_kwargs.get("inputs_embeds", None)
            )

            def _repeat_interleave_samples(x, lengths, repeat_times):
                samples = torch.split(x, lengths)
                repeat_args = [repeat_times] + [1] * (x.dim() - 1)
                result = torch.cat([sample.repeat(*repeat_args) for sample in samples], dim=0)
                return result

            for key in dict_to_expand:
                if key == "pixel_values":
                    # split images into samples
                    samples = torch.split(image_grid_thw, list(image_nums))
                    # compute the sequence length of images for each sample
                    lengths = [torch.prod(sample, dim=1).sum() for sample in samples]
                    dict_to_expand[key] = _repeat_interleave_samples(
                        dict_to_expand[key], lengths=lengths, repeat_times=expand_size
                    )
                elif key == "image_grid_thw":
                    # get the num of images for each sample
                    lengths = list(image_nums)
                    dict_to_expand[key] = _repeat_interleave_samples(
                        dict_to_expand[key], lengths=lengths, repeat_times=expand_size
                    )
                elif key == "pixel_values_videos":
                    samples = torch.split(video_grid_thw, list(video_nums))
                    lengths = [torch.prod(sample, dim=1).sum() for sample in samples]
                    dict_to_expand[key] = _repeat_interleave_samples(
                        dict_to_expand[key], lengths=lengths, repeat_times=expand_size
                    )
                elif key == "video_grid_thw":
                    lengths = list(video_nums)
                    dict_to_expand[key] = _repeat_interleave_samples(
                        dict_to_expand[key], lengths=lengths, repeat_times=expand_size
                    )
                elif key == "second_per_grid_ts":
                    dict_to_expand[key] = _repeat_interleave_samples(
                        dict_to_expand[key], lengths=list(video_nums), repeat_times=expand_size
                    )
            return dict_to_expand

        def _expand_dict_for_generation(dict_to_expand):
            for key in dict_to_expand:
                if (
                    key != "cache_position"
                    and dict_to_expand[key] is not None
                    and isinstance(dict_to_expand[key], torch.Tensor)
                    and key not in visual_keys
                ):
                    dict_to_expand[key] = dict_to_expand[key].repeat_interleave(expand_size, dim=0)
            return dict_to_expand

        model_kwargs = _expand_dict_for_generation_visual(model_kwargs)

        if input_ids is not None:
            input_ids = input_ids.repeat_interleave(expand_size, dim=0)

        model_kwargs = _expand_dict_for_generation(model_kwargs)

        if is_encoder_decoder:
            if model_kwargs.get("encoder_outputs") is None:
                raise ValueError("If `is_encoder_decoder` is True, make sure that `encoder_outputs` is defined.")
            model_kwargs["encoder_outputs"] = _expand_dict_for_generation(model_kwargs["encoder_outputs"])

        return input_ids, model_kwargs


__all__ = ["Qwen2_5_VLForConditionalGeneration", "Qwen2_5_VLModel", "Qwen2_5_VLPreTrainedModel", "Qwen2_5_VLTextModel"]


# ─────────────────────────────────────────────────────────────────────────────
# Vision processing utility functions (formerly qwen_vl_utils.py)
# ─────────────────────────────────────────────────────────────────────────────

def round_by_factor(number: int, factor: int) -> int:
    """Returns the closest integer to 'number' that is divisible by 'factor'."""
    return round(number / factor) * factor


def ceil_by_factor(number: int, factor: int) -> int:
    """Returns the smallest integer greater than or equal to 'number' that is divisible by 'factor'."""
    return math.ceil(number / factor) * factor


def floor_by_factor(number: int, factor: int) -> int:
    """Returns the largest integer less than or equal to 'number' that is divisible by 'factor'."""
    return math.floor(number / factor) * factor


def smart_resize(height: int, width: int, factor: int, min_pixels: Optional[int] = None, max_pixels: Optional[int] = None) -> Tuple[int, int]:
    """
    Rescales the image so that the following conditions are met:

    1. Both dimensions (height and width) are divisible by 'factor'.
    2. The total number of pixels is within the range ['min_pixels', 'max_pixels'].
    3. The aspect ratio of the image is maintained as closely as possible.
    """
    max_pixels = max_pixels if max_pixels is not None else (IMAGE_MAX_TOKEN_NUM * factor ** 2)
    min_pixels = min_pixels if min_pixels is not None else (IMAGE_MIN_TOKEN_NUM * factor ** 2)
    assert max_pixels >= min_pixels, "The max_pixels of image must be greater than or equal to min_pixels."
    if max(height, width) / min(height, width) > MAX_RATIO:
        raise ValueError(
            f"absolute aspect ratio must be smaller than {MAX_RATIO}, got {max(height, width) / min(height, width)}"
        )
    h_bar = max(factor, round_by_factor(height, factor))
    w_bar = max(factor, round_by_factor(width, factor))
    if h_bar * w_bar > max_pixels:
        beta = math.sqrt((height * width) / max_pixels)
        h_bar = floor_by_factor(height / beta, factor)
        w_bar = floor_by_factor(width / beta, factor)
    elif h_bar * w_bar < min_pixels:
        beta = math.sqrt(min_pixels / (height * width))
        h_bar = ceil_by_factor(height * beta, factor)
        w_bar = ceil_by_factor(width * beta, factor)
    return h_bar, w_bar


def to_rgb(pil_image: Image.Image) -> Image.Image:
      if pil_image.mode == 'RGBA':
          white_background = Image.new("RGB", pil_image.size, (255, 255, 255))
          white_background.paste(pil_image, mask=pil_image.split()[3])  # Use alpha channel as mask
          return white_background
      else:
          return pil_image.convert("RGB")


def fetch_image(ele: Dict[str, Union[str, Image.Image]], image_patch_size: int = 14) -> Image.Image:
    if "image" in ele:
        image = ele["image"]
    else:
        image = ele["image_url"]

    image_obj = None
    patch_factor = int(image_patch_size * SPATIAL_MERGE_SIZE)
    if isinstance(image, Image.Image):
        image_obj = image
    elif image.startswith("http://") or image.startswith("https://"):
        with requests.get(image, stream=True) as response:
            response.raise_for_status()
            with BytesIO(response.content) as bio:
                image_obj = copy.deepcopy(Image.open(bio))
    elif image.startswith("file://"):
        image_obj = Image.open(image[7:])
    elif image.startswith("data:image"):
        if "base64," in image:
            _, base64_data = image.split("base64,", 1)
            data = base64.b64decode(base64_data)
            with BytesIO(data) as bio:
                image_obj = copy.deepcopy(Image.open(bio))
    else:
        image_obj = Image.open(image)
    if image_obj is None:
        raise ValueError(f"Unrecognized image input, support local path, http url, base64 and PIL.Image, got {image}")
    image = to_rgb(image_obj)

    ## resize
    if "resized_height" in ele and "resized_width" in ele:
        resized_height, resized_width = smart_resize(
            ele["resized_height"],
            ele["resized_width"],
            factor=patch_factor,
        )
    else:
        width, height = image.size
        min_pixels = ele.get("min_pixels", IMAGE_MIN_TOKEN_NUM * patch_factor ** 2)
        max_pixels = ele.get("max_pixels", IMAGE_MAX_TOKEN_NUM * patch_factor ** 2)
        resized_height, resized_width = smart_resize(
            height,
            width,
            factor=patch_factor,
            min_pixels=min_pixels,
            max_pixels=max_pixels,
        )
    image = image.resize((resized_width, resized_height))
    return image


def smart_nframes(
    ele: Dict[str, Any],
    total_frames: int,
    video_fps: Union[int, float],
) -> int:
    """calculate the number of frames for video used for model inputs.

    Args:
        ele (dict): a dict contains the configuration of video.
            support either `fps` or `nframes`:
                - nframes: the number of frames to extract for model inputs.
                - fps: the fps to extract frames for model inputs.
                    - min_frames: the minimum number of frames of the video, only used when fps is provided.
                    - max_frames: the maximum number of frames of the video, only used when fps is provided.
        total_frames (int): the original total number of frames of the video.
        video_fps (int | float): the original fps of the video.

    Raises:
        ValueError: nframes should in interval [FRAME_FACTOR, total_frames].

    Returns:
        int: the number of frames for video used for model inputs.
    """
    assert not ("fps" in ele and "nframes" in ele), "Only accept either `fps` or `nframes`"
    if "nframes" in ele:
        nframes = round_by_factor(ele["nframes"], FRAME_FACTOR)
    else:
        fps = ele.get("fps", FPS)
        min_frames = ceil_by_factor(ele.get("min_frames", FPS_MIN_FRAMES), FRAME_FACTOR)
        max_frames = floor_by_factor(ele.get("max_frames", min(FPS_MAX_FRAMES, total_frames)), FRAME_FACTOR)
        nframes = total_frames / video_fps * fps
        if nframes > total_frames:
            logger.warning(f"smart_nframes: nframes[{nframes}] > total_frames[{total_frames}]")
        nframes = min(min(max(nframes, min_frames), max_frames), total_frames)
        nframes = floor_by_factor(nframes, FRAME_FACTOR)
    if not (FRAME_FACTOR <= nframes and nframes <= total_frames):
        raise ValueError(f"nframes should in interval [{FRAME_FACTOR}, {total_frames}], but got {nframes}.")
    return nframes


def _read_video_torchvision(
    ele: Dict[str, Any],
) -> Tuple[torch.Tensor, float]:
    """read video using torchvision.io.read_video

    Args:
        ele (dict): a dict contains the configuration of video.
        support keys:
            - video: the path of video. support "file://", "http://", "https://" and local path.
            - video_start: the start time of video.
            - video_end: the end time of video.
    Returns:
        torch.Tensor: the video tensor with shape (T, C, H, W).
    """
    video_path = ele["video"]
    if version.parse(torchvision.__version__) < version.parse("0.19.0"):
        if "http://" in video_path or "https://" in video_path:
            warnings.warn("torchvision < 0.19.0 does not support http/https video path, please upgrade to 0.19.0.")
        if "file://" in video_path:
            video_path = video_path[7:]
    st = time.time()
    video, audio, info = io.read_video(
        video_path,
        start_pts=ele.get("video_start", 0.0),
        end_pts=ele.get("video_end", None),
        pts_unit="sec",
        output_format="TCHW",
    )
    total_frames, video_fps = video.size(0), info["video_fps"]
    logger.info(f"torchvision:  {video_path=}, {total_frames=}, {video_fps=}, time={time.time() - st:.3f}s")
    nframes = smart_nframes(ele, total_frames=total_frames, video_fps=video_fps)
    idx = torch.linspace(0, total_frames - 1, nframes).round().long()
    sample_fps = nframes / max(total_frames, 1e-6) * video_fps
    video = video[idx]

    video_metadata = dict(
        fps=video_fps,
        frames_indices=idx,
        total_num_frames=total_frames,
        video_backend="torchvision",
    )
    return video, video_metadata, sample_fps


def is_decord_available() -> bool:
    import importlib.util

    return importlib.util.find_spec("decord") is not None


def calculate_video_frame_range(
    ele: Dict[str, Any],
    total_frames: int,
    video_fps: float,
) -> Tuple[int, int, int]:
    """
    Calculate the start and end frame indices based on the given time range.

    Args:
        ele (dict): A dictionary containing optional 'video_start' and 'video_end' keys (in seconds).
        total_frames (int): Total number of frames in the video.
        video_fps (float): Frames per second of the video.

    Returns:
        tuple: A tuple containing (start_frame, end_frame, frame_count).

    Raises:
        ValueError: If input parameters are invalid or the time range is inconsistent.
    """
    # Validate essential parameters
    if video_fps <= 0:
        raise ValueError("video_fps must be a positive number")
    if total_frames <= 0:
        raise ValueError("total_frames must be a positive integer")

    # Get start and end time in seconds
    video_start = ele.get("video_start", None)
    video_end = ele.get("video_end", None)
    if video_start is None and video_end is None:
        return 0, total_frames - 1, total_frames

    max_duration = total_frames / video_fps
    # Process start frame
    if video_start is not None:
        video_start_clamped = max(0.0, min(video_start, max_duration))
        start_frame = math.ceil(video_start_clamped * video_fps)
    else:
        start_frame = 0
    # Process end frame
    if video_end is not None:
        video_end_clamped = max(0.0, min(video_end, max_duration))
        end_frame = math.floor(video_end_clamped * video_fps)
        end_frame = min(end_frame, total_frames - 1)
    else:
        end_frame = total_frames - 1

    # Validate frame order
    if start_frame >= end_frame:
        raise ValueError(
            f"Invalid time range: Start frame {start_frame} (at {video_start_clamped if video_start is not None else 0}s) "
            f"exceeds end frame {end_frame} (at {video_end_clamped if video_end is not None else max_duration}s). "
            f"Video duration: {max_duration:.2f}s ({total_frames} frames @ {video_fps}fps)"
        )

    logger.info(f"calculate video frame range: {start_frame=}, {end_frame=}, {total_frames=} from {video_start=}, {video_end=}, {video_fps=:.3f}")
    return start_frame, end_frame, end_frame - start_frame + 1


def _read_video_decord(
    ele: Dict[str, Any],
) -> Tuple[torch.Tensor, float]:
    """read video using decord.VideoReader

    Args:
        ele (dict): a dict contains the configuration of video.
        support keys:
            - video: the path of video. support "file://", "http://", "https://" and local path.
            - video_start: the start time of video.
            - video_end: the end time of video.
    Returns:
        torch.Tensor: the video tensor with shape (T, C, H, W).
    """
    import decord
    video_path = ele["video"]
    st = time.time()
    vr = decord.VideoReader(video_path)
    total_frames, video_fps = len(vr), vr.get_avg_fps()
    start_frame, end_frame, total_frames = calculate_video_frame_range(
        ele,
        total_frames,
        video_fps,
    )
    nframes = smart_nframes(ele, total_frames=total_frames, video_fps=video_fps)
    idx = torch.linspace(start_frame, end_frame, nframes).round().long().tolist()
    video = vr.get_batch(idx).asnumpy()
    video = torch.tensor(video).permute(0, 3, 1, 2)  # Convert to TCHW format
    logger.info(f"decord:  {video_path=}, {total_frames=}, {video_fps=}, time={time.time() - st:.3f}s")
    sample_fps = nframes / max(total_frames, 1e-6) * video_fps

    video_metadata = dict(
        fps=video_fps,
        frames_indices=idx,
        total_num_frames=total_frames,
        video_backend="decord",
    )
    return video, video_metadata, sample_fps


def is_torchcodec_available() -> bool:
    import importlib.util

    return importlib.util.find_spec("torchcodec") is not None


def _read_video_torchcodec(
    ele: Dict[str, Any],
) -> Tuple[torch.Tensor, float]:
    """read video using torchcodec.decoders.VideoDecoder

    Args:
        ele (dict): a dict contains the configuration of video.
        support keys:
            - video: the path of video. support "file://", "http://", "https://" and local path.
            - video_start: the start time of video.
            - video_end: the end time of video.
    Returns:
        torch.Tensor: the video tensor with shape (T, C, H, W).
    """
    from torchcodec.decoders import VideoDecoder
    TORCHCODEC_NUM_THREADS = int(os.environ.get('TORCHCODEC_NUM_THREADS', 8))
    logger.info(f"set TORCHCODEC_NUM_THREADS: {TORCHCODEC_NUM_THREADS}")
    video_path = ele["video"]
    st = time.time()
    decoder = VideoDecoder(video_path, num_ffmpeg_threads=TORCHCODEC_NUM_THREADS)
    video_fps = decoder.metadata.average_fps
    total_frames = decoder.metadata.num_frames
    start_frame, end_frame, total_frames = calculate_video_frame_range(
        ele,
        total_frames,
        video_fps,
    )
    nframes = smart_nframes(ele, total_frames=total_frames, video_fps=video_fps)
    idx = torch.linspace(start_frame, end_frame, nframes).round().long().tolist()
    sample_fps = nframes / max(total_frames, 1e-6) * video_fps
    video = decoder.get_frames_at(indices=idx).data
    logger.info(f"torchcodec:  {video_path=}, {total_frames=}, {video_fps=}, time={time.time() - st:.3f}s")

    video_metadata = dict(
        fps=video_fps,
        frames_indices=idx,
        total_num_frames=total_frames,
        video_backend="torchcodec",
    )
    return video, video_metadata, sample_fps


VIDEO_READER_BACKENDS = {
    "decord": _read_video_decord,
    "torchvision": _read_video_torchvision,
    "torchcodec": _read_video_torchcodec,
}

FORCE_QWENVL_VIDEO_READER = os.getenv("FORCE_QWENVL_VIDEO_READER", None)


@lru_cache(maxsize=1)
def get_video_reader_backend() -> str:
    if FORCE_QWENVL_VIDEO_READER is not None:
        video_reader_backend = FORCE_QWENVL_VIDEO_READER
    elif is_torchcodec_available():
        video_reader_backend = "torchcodec"
    elif is_decord_available():
        video_reader_backend = "decord"
    else:
        video_reader_backend = "torchvision"
    print(f"qwen-vl-utils using {video_reader_backend} to read video.", file=sys.stderr)
    return video_reader_backend


def fetch_video(ele: Dict[str, Any], image_patch_size: int = 14, return_video_sample_fps: bool = False,
                return_video_metadata: bool = False) -> Union[torch.Tensor, List[Image.Image]]:
    image_factor = image_patch_size * SPATIAL_MERGE_SIZE
    VIDEO_FRAME_MIN_PIXELS = VIDEO_MIN_TOKEN_NUM * image_factor * image_factor
    VIDEO_FRAME_MAX_PIXELS = VIDEO_MAX_TOKEN_NUM * image_factor * image_factor
    if isinstance(ele["video"], str):
        video_reader_backend = get_video_reader_backend()
        try:
            video, video_metadata, sample_fps = VIDEO_READER_BACKENDS[video_reader_backend](ele)
        except Exception as e:
            logger.warning(f"video_reader_backend {video_reader_backend} error, use torchvision as default, msg: {e}")
            video, video_metadata, sample_fps = VIDEO_READER_BACKENDS["torchvision"](ele)
    else:
        # The input is a list of frames
        # assert isinstance(ele["video"], (list, tuple))
        process_info = ele.copy()
        process_info.pop("type", None)
        process_info.pop("video", None)
        if len(ele["video"]) > FPS_MAX_FRAMES:
            nframes = FPS_MAX_FRAMES
        else:
            nframes = len(ele["video"])
        if len(ele["video"]) > nframes:
            idx = torch.linspace(0, len(ele["video"]) - 1, nframes).round().long()
            ele["video"] = [ele["video"][i] for i in idx]
        # use ThreadPoolExecutor to parallel process frames
        max_workers = min(MAX_NUM_WORKERS_FETCH_VIDEO, len(ele["video"]))
        with ThreadPoolExecutor(max_workers=max_workers) as executor:
            futures = [
                executor.submit(fetch_image, {"image": video_element, **process_info}, image_factor)
                for video_element in ele["video"]
            ]
            image_list = [future.result() for future in futures]

        # nframes = ceil_by_factor(len(image_list), FRAME_FACTOR)
        # if len(image_list) < nframes:
        #     image_list.extend([image_list[-1]] * (nframes - len(image_list)))
        sample_fps = ele.get("sample_fps", 10.0)
        video = torch.stack([
            torch.from_numpy(np.array(image).transpose(2, 0, 1))
            for image in image_list
        ])

        # fake video metadata
        raw_fps = process_info.pop("raw_fps", sample_fps)
        video_metadata = dict(
            fps=raw_fps,
            frames_indices=[i for i in range(len(video))],
            total_num_frames=(nframes / sample_fps) * raw_fps,
        )

    nframes, _, height, width = video.shape
    min_pixels = ele.get("min_pixels", VIDEO_FRAME_MIN_PIXELS)
    total_pixels = ele.get("total_pixels", MODEL_SEQ_LEN * image_factor * image_factor * 0.9)
    max_pixels = max(min(VIDEO_FRAME_MAX_PIXELS, total_pixels / nframes * FRAME_FACTOR), int(min_pixels * 1.05))
    max_pixels_supposed = ele.get("max_pixels", max_pixels)
    if max_pixels_supposed > max_pixels:
        logger.warning(f"The given max_pixels[{max_pixels_supposed}] exceeds limit[{max_pixels}].")
    max_pixels = min(max_pixels_supposed, max_pixels)
    if "resized_height" in ele and "resized_width" in ele:
        resized_height, resized_width = smart_resize(
            ele["resized_height"],
            ele["resized_width"],
            factor=image_factor,
        )
    else:
        resized_height, resized_width = smart_resize(
            height,
            width,
            factor=image_factor,
            min_pixels=min_pixels,
            max_pixels=max_pixels,
        )
    video = transforms.functional.resize(
        video,
        [resized_height, resized_width],
        interpolation=InterpolationMode.BICUBIC,
        antialias=True,
    ).float()

    final_video = (video, video_metadata) if return_video_metadata else video
    if return_video_sample_fps:
        return final_video, sample_fps
    return final_video


def extract_vision_info(conversations: Union[List[Dict[str, Any]], List[List[Dict[str, Any]]]]) -> List[Dict[str, Any]]:
    vision_infos = []
    if isinstance(conversations[0], dict):
        conversations = [conversations]
    for conversation in conversations:
        for message in conversation:
            if isinstance(message["content"], list):
                for ele in message["content"]:
                    if (
                        "image" in ele
                        or "image_url" in ele
                        or "video" in ele
                        or ele.get("type", "text") in ("image", "image_url", "video")
                    ):
                        vision_infos.append(ele)
    return vision_infos


def process_vision_info(
    conversations: Union[List[Dict[str, Any]], List[List[Dict[str, Any]]]],
    return_video_kwargs: bool = False,
    return_video_metadata: bool = False,
    image_patch_size: int = 14,
) -> Tuple[Optional[List[Image.Image]], Optional[List[Union[torch.Tensor, List[Image.Image]]]], Optional[Dict[str, Any]]]:

    vision_infos = extract_vision_info(conversations)
    ## Read images or videos
    image_inputs = []
    video_inputs = []
    video_sample_fps_list = []
    for vision_info in vision_infos:
        if "image" in vision_info or "image_url" in vision_info:
            image_inputs.append(fetch_image(vision_info, image_patch_size=image_patch_size))
        elif "video" in vision_info:
            video_input, video_sample_fps = fetch_video(vision_info, return_video_sample_fps=True,
                        image_patch_size=image_patch_size, return_video_metadata=return_video_metadata)
            video_sample_fps_list.append(video_sample_fps)
            video_inputs.append(video_input)
        else:
            raise ValueError("image, image_url or video should in content.")
    if len(image_inputs) == 0:
        image_inputs = None
    if len(video_inputs) == 0:
        video_inputs = None

    video_kwargs = {'do_sample_frames': False}
    if not return_video_metadata: # BC for qwen2.5vl
        video_kwargs.update({'fps': video_sample_fps_list})

    if return_video_kwargs:
        return image_inputs, video_inputs, video_kwargs
    return image_inputs, video_inputs

# ─────────────────────────────────────────────────────────────────────────────
# MLLM Encoder (diffusers-compatible wrapper)
# ─────────────────────────────────────────────────────────────────────────────

class MLLMEncoder(ModelMixin, ConfigMixin):
    """
    Multimodal LLM encoder for KiwiEdit based on Qwen2.5-VL.
    Processes text + source video/image + optional reference image
    through learnable query mechanism with dual connector architecture.
    """

    @register_to_config
    def __init__(
        self,
        mllm_model_path: str = "Qwen/Qwen2.5-VL-3B-Instruct",
        dit_dim: int = 3072,
        hidden_size: int = 2048,
        num_image_queries: int = 256,
        num_video_queries: int = 512,
        num_ref_queries: int = 768,
        max_object_token: int = 768,
        max_frames: int = 16,
        max_pixels_per_frame: int = 262144,
    ):
        super().__init__()
        self._mllm_model_path = mllm_model_path
        self.num_image_queries = num_image_queries
        self.num_video_queries = num_video_queries
        self.num_ref_queries = num_ref_queries
        self.max_object_token = max_object_token
        self.max_frames = max_frames
        self.max_pixels_per_frame = max_pixels_per_frame

        # Learnable query vectors
        self.image_queries = nn.Parameter(
            torch.randn(num_image_queries, hidden_size) * 0.02
        )
        self.video_queries = nn.Parameter(
            torch.randn(num_video_queries, hidden_size) * 0.02
        )
        self.ref_queries = nn.Parameter(
            torch.randn(num_ref_queries, hidden_size) * 0.02
        )

        # Connector MLP: MLLM hidden → DiT dim
        self.connector = nn.Sequential(
            nn.Linear(hidden_size, dit_dim),
            nn.GELU(approximate="tanh"),
            nn.Linear(dit_dim, dit_dim),
        )
        nn.init.zeros_(self.connector[2].weight)
        nn.init.zeros_(self.connector[2].bias)

        # Ref connector MLP (separate from main connector)
        self.ref_connector = nn.Sequential(
            nn.Linear(hidden_size, dit_dim),
            nn.GELU(approximate="tanh"),
            nn.Linear(dit_dim, dit_dim),
        )
        nn.init.zeros_(self.ref_connector[2].weight)
        nn.init.zeros_(self.ref_connector[2].bias)

        # Qwen VL model and processor (loaded lazily)
        self.qwen_model = None
        self.processor = None
        self.system_prompt = (
            "You will be given an image and instruction. "
            "Please describe the content of the image in detail "
            "based on instruction in your own words."
        )

    def _resolve_qwen_path(self):
        """Resolve the path where Qwen model files live.

        With the split layout, Qwen weights live alongside the MLLMEncoder config.json,
        while processor/tokenizer assets live under a sibling processor/ folder.
        config._name_or_path points to that directory (local) or is an HF repo ID.
        """
        name_or_path = getattr(getattr(self, "config", None), "_name_or_path", None) or ""
        module_dir = os.path.dirname(os.path.abspath(__file__))

        def _is_qwen_dir(path: str) -> bool:
            if not path or not os.path.isdir(path):
                return False
            return any(
                os.path.isfile(os.path.join(path, fname))
                for fname in (
                    "qwen_config.json",
                    "model.safetensors.index.json",
                    "model-00001-of-00002.safetensors",
                )
            )

        # "." means flat layout: Qwen files live alongside the MLLMEncoder config
        if self._mllm_model_path == ".":
            # 1) Already-resolved local directories.
            for candidate in (
                name_or_path,
                module_dir,
                os.path.join(module_dir, "mllm_encoder"),
            ):
                if _is_qwen_dir(candidate):
                    return candidate

            # 2) Diffusers may keep the local path on different attributes.
            for attr in ("_pretrained_model_name_or_path", "name_or_path"):
                local = getattr(self, attr, None)
                if isinstance(local, str):
                    if _is_qwen_dir(local):
                        return local
                    local_sub = os.path.join(local, "mllm_encoder")
                    if _is_qwen_dir(local_sub):
                        return local_sub

            # 3) name_or_path is an HF repo ID — try cache first, then network.
            if name_or_path and "/" in name_or_path:
                try:
                    from huggingface_hub import snapshot_download
                    cached = snapshot_download(
                        name_or_path,
                        allow_patterns=["mllm_encoder/*"],
                        local_files_only=True,
                    )
                    local = os.path.join(cached, "mllm_encoder")
                    if _is_qwen_dir(local):
                        return local
                except Exception:
                    pass
                try:
                    from huggingface_hub import snapshot_download

                    cached = snapshot_download(name_or_path, allow_patterns=["mllm_encoder/*"])
                    local = os.path.join(cached, "mllm_encoder")
                    if _is_qwen_dir(local):
                        return local
                except Exception:
                    pass

        # If mllm_model_path is an absolute path, use as-is
        if os.path.isabs(self._mllm_model_path) and os.path.isdir(self._mllm_model_path):
            return self._mllm_model_path

        # Check relative to config._name_or_path
        if name_or_path and os.path.isdir(name_or_path):
            # Flat layout: check if qwen_config.json exists in config._name_or_path
            if os.path.isfile(os.path.join(name_or_path, "qwen_config.json")):
                return name_or_path
            # Legacy: check for subdirectory (e.g. qwen_model/)
            sub = os.path.join(name_or_path, self._mllm_model_path)
            if os.path.isdir(sub):
                return sub

        # Fallback: treat as HuggingFace model ID
        return self._mllm_model_path

    def _resolve_processor_path(self) -> Optional[str]:
        """Resolve the path where processor/tokenizer files live."""
        name_or_path = getattr(getattr(self, "config", None), "_name_or_path", None) or ""
        module_dir = os.path.dirname(os.path.abspath(__file__))

        def _is_processor_dir(path: str) -> bool:
            if not path or not os.path.isdir(path):
                return False
            return any(
                os.path.isfile(os.path.join(path, fname))
                for fname in (
                    "tokenizer_config.json",
                    "tokenizer.json",
                    "preprocessor_config.json",
                    "chat_template.jinja",
                    "vocab.json",
                )
            )

        processor_hint = getattr(self, "_processor_path", None)
        if isinstance(processor_hint, str) and _is_processor_dir(processor_hint):
            return processor_hint

        for candidate in (
            os.path.join(module_dir, "processor"),
            os.path.join(os.path.dirname(module_dir), "processor"),
        ):
            candidate = os.path.abspath(candidate)
            if _is_processor_dir(candidate):
                return candidate

        if name_or_path and os.path.isdir(name_or_path):
            for candidate in (
                os.path.join(name_or_path, "processor"),
                os.path.join(os.path.dirname(name_or_path), "processor"),
            ):
                if _is_processor_dir(candidate):
                    return candidate

        for attr in ("_pretrained_model_name_or_path", "name_or_path"):
            local = getattr(self, attr, None)
            if isinstance(local, str) and os.path.isdir(local):
                for candidate in (
                    os.path.join(local, "processor"),
                    os.path.join(os.path.dirname(local), "processor"),
                ):
                    if _is_processor_dir(candidate):
                        return candidate

        if name_or_path and "/" in name_or_path:
            try:
                from huggingface_hub import snapshot_download
                cached = snapshot_download(
                    name_or_path,
                    allow_patterns=["processor/*"],
                    local_files_only=True,
                )
                local = os.path.join(cached, "processor")
                if _is_processor_dir(local):
                    return local
            except Exception:
                pass
            try:
                from huggingface_hub import snapshot_download
                cached = snapshot_download(name_or_path, allow_patterns=["processor/*"])
                local = os.path.join(cached, "processor")
                if _is_processor_dir(local):
                    return local
            except Exception:
                pass

        return None

    def load_qwen_model(self, device="cuda", dtype=torch.bfloat16):
        """Load the Qwen VL model and processor. Call this after from_pretrained."""
        from transformers import AutoProcessor, AutoConfig

        qwen_path = self._resolve_qwen_path()
        processor_path = self._resolve_processor_path()
        device_map = str(device) if isinstance(device, torch.device) else device
        if isinstance(device_map, str) and device_map not in {
            "auto",
            "balanced",
            "balanced_low_0",
            "sequential",
            "cpu",
        }:
            # Transformers expects a mapping for explicit single-device placement.
            device_map = {"": device_map}

        # Flat layout uses qwen_config.json to avoid conflict with diffusers config.json
        qwen_config_file = os.path.join(qwen_path, "qwen_config.json")
        if not os.path.isfile(qwen_config_file) and processor_path:
            qwen_config_file = os.path.join(processor_path, "qwen_config.json")
        if os.path.isfile(qwen_config_file):
            qwen_config = AutoConfig.from_pretrained(qwen_config_file)
        else:
            extra = f" or {processor_path}" if processor_path else ""
            raise FileNotFoundError(
                f"qwen_config.json is missing under {qwen_path}{extra}. "
                "Ensure the processor folder includes the Qwen config."
            )

        self.qwen_model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
            qwen_path, config=qwen_config, torch_dtype=dtype, device_map=device_map
        )
        self.qwen_model.eval()
        hidden_size = self.qwen_model.config.hidden_size
        # Set query counts on the Qwen model
        self.qwen_model.model.num_image_queries = self.num_image_queries
        self.qwen_model.num_image_queries = self.num_image_queries
        self.qwen_model.model.num_video_queries = self.num_video_queries
        self.qwen_model.num_video_queries = self.num_video_queries
        self.qwen_model.model.num_ref_queries = self.num_ref_queries

        if self.processor is None:
            processor_base = processor_path or qwen_path
            for fname in ("tokenizer_config.json", "tokenizer.json", "preprocessor_config.json"):
                if not os.path.isfile(os.path.join(processor_base, fname)):
                    raise FileNotFoundError(
                        f"Processor asset {fname} is missing under {processor_base}. "
                        "Ensure the processor folder includes the Qwen processor files."
                    )
            self.processor = AutoProcessor.from_pretrained(processor_base)
        return hidden_size

    def _ensure_qwen_loaded(self):
        if self.qwen_model is None:
            device = next(self.parameters()).device
            dtype = next(self.parameters()).dtype
            self.load_qwen_model(device=device, dtype=dtype)

    @torch.no_grad()
    def forward(
        self,
        instruction: str,
        src_image: Optional[List[Image.Image]] = None,
        src_video: Optional[List[Image.Image]] = None,
        ref_image: Optional[List[Image.Image]] = None,
    ) -> torch.Tensor:
        """
        Encode text + visual inputs through MLLM with learnable queries.

        Args:
            instruction: Text editing instruction.
            src_image: Source image(s) for image editing mode.
            src_video: Source video frames for video editing mode.
            ref_image: Reference image(s) for reference-guided editing.

        Returns:
            Context embeddings of shape [1, seq_len, dit_dim].
        """
        self._ensure_qwen_loaded()

        is_video = src_video is not None
        system_prompt = {
            "role": "system",
            "content": [{"type": "text", "text": self.system_prompt}],
        }

        # Build messages based on mode
        if ref_image:
            # Reference-guided video editing
            instruction_text = instruction + " Use the reference input from last frame."
            num_queries = self.num_ref_queries
            input_query = self.ref_queries.to(self.qwen_model.device)
            video_data = {
                "type": "video",
                "video": src_video,
                "max_frames": self.max_frames,
            }
            if self.max_pixels_per_frame:
                video_data["max_pixels"] = self.max_pixels_per_frame
            messages = [
                system_prompt,
                {
                    "role": "user",
                    "content": [
                        video_data,
                        {"type": "text", "text": instruction_text},
                        {
                            "type": "image",
                            "image": ref_image[0],
                            "max_pixels": 28 * 28 * self.max_object_token,
                        },
                        {
                            "type": "text",
                            "text": "<|object_ref_start|>" * num_queries,
                        },
                    ],
                },
            ]
        elif is_video:
            # Instruction-only video editing
            num_queries = self.num_video_queries
            input_query = self.video_queries.to(self.qwen_model.device)
            video_data = {
                "type": "video",
                "video": src_video,
                "max_frames": self.max_frames,
            }
            if self.max_pixels_per_frame:
                video_data["max_pixels"] = self.max_pixels_per_frame
            messages = [
                system_prompt,
                {
                    "role": "user",
                    "content": [
                        video_data,
                        {"type": "text", "text": instruction},
                        {
                            "type": "text",
                            "text": "<|object_ref_start|>" * num_queries,
                        },
                    ],
                },
            ]
        else:
            # Image editing
            num_queries = self.num_image_queries
            input_query = self.image_queries.to(self.qwen_model.device)
            messages = [
                system_prompt,
                {
                    "role": "user",
                    "content": [
                        {"type": "image", "image": src_image[0]},
                        {"type": "text", "text": instruction},
                        {
                            "type": "text",
                            "text": "<|object_ref_start|>" * num_queries,
                        },
                    ],
                },
            ]

        # Process through Qwen
        text = self.processor.apply_chat_template(
            messages, tokenize=False, add_generation_prompt=False
        )
        image_inputs, video_inputs = process_vision_info(messages)
        inputs = self.processor(
            text=[text],
            images=image_inputs,
            videos=video_inputs,
            padding=True,
            return_tensors="pt",
        )
        inputs = inputs.to(self.qwen_model.device, dtype=torch.bfloat16)

        outputs = self.qwen_model(
            **inputs,
            output_attentions=False,
            output_hidden_states=True,
            return_dict=True,
            learnable_query=input_query,
        )
        hidden_states = outputs.hidden_states[-1]
        learnable_query_features = hidden_states[:, -input_query.shape[0] :, :]
        learnable_query_features = self.connector(learnable_query_features)

        # Extract ref image features if in ref mode
        if ref_image:
            vision_start_id = self.processor.tokenizer.convert_tokens_to_ids(
                "<|vision_start|>"
            )
            vision_end_id = self.processor.tokenizer.convert_tokens_to_ids(
                "<|vision_end|>"
            )
            input_ids = inputs.input_ids[0]
            vision_start_indices = (input_ids == vision_start_id).nonzero(
                as_tuple=True
            )[-1]
            if len(vision_start_indices) > 0:
                last_vision_start = vision_start_indices[-1]
                remaining_ids = input_ids[last_vision_start:]
                end_relative_idx = (remaining_ids == vision_end_id).nonzero(
                    as_tuple=True
                )[-1]
                if len(end_relative_idx) > 0:
                    last_vision_end = last_vision_start + end_relative_idx[0]
                    ref_image_features = hidden_states[
                        :, last_vision_start + 1 : last_vision_end, :
                    ]
                    ref_image_features = self.ref_connector(ref_image_features)
                    learnable_query_features = torch.cat(
                        [ref_image_features, learnable_query_features], dim=1
                    )

        return learnable_query_features